Abdallah Attia, Jihun Hamm, Mahmoud A AbdAlnaeem, Zhengming Ding, Michael O'Rorke, Joseph Dillon, Mary Maluccio, Nicholas Skill, Kristen Limbach

Methods. Forty-four patients with histologically confirmed PNET who underwent contrast-enhanced preoperative CT and surgical resection at two academic centers were analyzed. Lesion and contralateral non-tumor-bearing pancreatic parenchyma regions of interest were revised in 3D Slicer by a board-certified pancreatic surgeon and verified intraoperatively against surgical pathology. PyRadiomics v3.0 features were extracted with IBSI-concordant settings. Parametric ComBat batch correction was applied across the two centers (biological-covariate balance verified beforehand), and Δ-radiomic features (lesion combat-pancreas combat) were computed for the 106 intensity/texture primitives. We constructed a panel of biology-informed hybrid signatures partitioned into a preoperative lesion-only family (Family A; seven signatures) and a preoperative Δ-radiomic family (Family B; three signatures). Candidate features were filtered through correlation clustering, baseline-adjusted likelihood-ratio testing with Benjamini-Hochberg FDR control, and 100-bootstrap stability selection. Three predictor blocks were compared per target with three classifiers each (Logistic Regression, Random Forest, Gradient Boosting): M0 (five-variable clinical baseline), MA (M0 + Family A), and MB (M0 + Family B). Discrimination was reported as AUC with bootstrap 95% CI; calibration was assessed using the Brier score and TRIPOD-recommended calibration intercept and slope; and cross-center generalization was evaluated with leave-one-center-out (LOCO) cross-validation. Univariable Cox regression with bootstrap and permutation inference was used for progression-free survival (PFS). 

Results. The cohort had 16 progression events and eight deaths (median follow-up was 38 months, IQR 14-59). Prespecified clinical-radiomic and Δ-radiomic signatures were associated with progression-free survival, including B2 = ΔBusyness � Ki-67 (HR 0.38, 95% CI 0.19-0.76, p = 0.006). For progression prediction, the Δ-radiomic model achieved the strongest discrimination, with a nested cross-validation AUC of 0.85 and leave-one-center-out AUC of 0.87. For higher-grade disease, radiomic models also demonstrated high discrimination, with AUCs up to 0.93. 

Conclusions. Radiomics-derived shape and texture features, especially when combined with clinical markers, may noninvasively identify aggressive PNET phenotypes and support preoperative risk stratification. Prospective validation in larger multicenter cohorts is warranted.